To all and sundry There's a thread spinning out in IBMTCP-L concerning trying to persuade a zIIP to do some work handling TCP
traffic supposedly exercising the "multiple write" feature of HiperSockets as described in the following two sections in the z/OS Communications Server (CS) IP Configuration Guide manual: <quote> 1.2.18.17.9 HiperSockets multiple write The HiperSockets multiple write facility moves multiple output data buffers in a single write operation. This facility might reduce CPU usage, and might provide a performance improvement for large outbound messages that are typically generated by traditional streaming workloads such as file transfer, and interactive web-based services workloads such as XML or SOAP. To enable the HiperSockets multiple write facility on all HiperSockets interfaces, including interfaces created for dynamic XCF, add the IQDMULTIWRITE parameter to the GLOBALCONFIG statement. For more information about the GLOBALCONFIG statement and the HiperSockets multiple write facility, see z/OS Communications Server: IP Configuration Reference. Restriction: HiperSockets multiple write is effective only on an IBM System z10 and when z/OS is not running as a guest in a z/VM environment. </quote> http://publibz.boulder.ibm.com/cgi- bin/bookmgr_OS390/BOOKS/F1A1B3A0/1.2.18.17.9 <quote> 1.2.18.17.10 HiperSockets multiple write assist with IBM zIIP When your system is an IBM System z10 and the HiperSockets multiple write facility is enabled, an additional assist for large outbound TCP messages is available through the IBM System z10 Integrated Information Processor and IBM System z9 Integrated Information Processor (zIIP). To enable HiperSockets write processing for large outbound TCP messages on available zIIPs, specify the ZIIP IQDIOMULTIWRITE parameter on the GLOBALCONFIG statement. When your system is an IBM System z10 that does not have zIIPs configured, you can use the zIIP HiperSockets multiple write facility to project the percentage of existing HiperSockets workload currently running on central processors that would be eligible to run on zIIPs, if zIIPs were available on the z/OS image. ... </quote> http://publibz.boulder.ibm.com/cgi- bin/bookmgr_OS390/BOOKS/F1A1B3A0/1.2.18.17.10 Apparently, if you manage to persuade the zIIP to earn its keep, you will see a non-zero percentage alongside the name of the main CS IP address space under "IIP" in RMF output similar to the following: <sample> . RMF V1R11 Processor Usage Line 1 of 94 . . Command ===> Scroll ===> CSR . . . . Samples: 120 System: CWD2 Date: 05/25/11 Time: 15.35.00 Range: 120 Sec . . . . Service --- Time on CP % --- ----- EAppl % ----- . . Jobname CX Class Total AAP IIP CP AAP IIP . . . . TCPIP SO SYSSTC 16.2 0.0 0.0 16.2 0.0 . . ... </sample> After some discussion - including the odd developer - the conclusion seems to be that the following are the necessary prerequisites: - z10 - an available zIIP - *not* a VM guest - using the TCP transport layer logic - socket send call size >= 32K - GLOBALCONFIG IQDMULTIWRITE - GLOBALCONFIG ZIIP IQDIOMULTIWRITE So my request to the list participants is to ask whether anyone has succeeded, in effect, in making that percentage use of the zIIP other than 0 and if so how. In particular, is that list of prerequisites correct and complete? In the case of successful use of the zIIP, what was the application involved. If Connect:Direct - using TCP and IP of course[1] - what were the parameters of the application which needed attention? If other, perhaps the same question applies so that others using that application might benefit eventually (in both the English and non-English meanings). - [1] It appears that output from commands to Connect:Direct even when using TCP and IP, contain the tokens "VTAM", "RUs" and "Rusize" which is (a) very lazy of the Connect:Direct developers and (b) guaranteed to cause confusion! No, the problem is not that an SNA flavour of Connect:Direct is being used over Enterprise Extender and hence that the transport protocol is UDP. That possibility has already been explored ad nauseam! - Chris Mason ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to [email protected] with the message: GET IBM-MAIN INFO Search the archives at http://bama.ua.edu/archives/ibm-main.html
